SandboxAQ Opens a Drug-Screening Model That Does Not Need Protein Structures

Many computational drug screens begin with a detailed 3D map of a target protein. SandboxAQ’s newly available AQPotency takes a different route: it ranks potential drug compounds without requiring that solved structure, aiming to make earlier screening possible for targets that cannot reliably enter a conventional structure-based workflow.

AQPotency predicts how potent a molecule may be against a biological target, expressed as a predicted pIC50 value. That score is meant to help researchers decide which candidates deserve laboratory testing; it does not establish that a compound will be effective or safe as a drug.

A screen before the wet lab

The structural requirement matters because many established virtual-screening workflows rely on a detailed protein representation. Experimental or computational structural models are not usable or reliable for every target. SandboxAQ says AQPotency can instead work from the molecule and target, allowing teams to rank candidates for programs that may not be able to start with conventional docking.

That leaves AQPotency positioned against both experimental and predicted protein maps, rather than only missing crystal structures. The company’s premise is that molecule-and-target inputs can still provide a useful ranking when a structure-based starting point is unavailable or unreliable.

What each AQPotency result is designed to provide

• A predicted potency score for a molecule-target comparison, expressed as pIC50.
• A confidence estimate and a signal of whether the target resembles the data range in which the model performs reliably.
• A reverse-search option that can rank proteins a molecule may affect, a workflow SandboxAQ positions for off-target screening or investigating an uncertain biological mechanism.

Those uncertainty signals are a meaningful part of the design. A ranking can be more useful when researchers can distinguish a prediction that fits the model’s known data range from one that reaches beyond it. But the confidence indicator is not itself experimental confirmation.

The product enters a field where vendors are tackling screening scale and structural uncertainty differently. Schrödinger’s virtual-screening service combines 3D docking with machine learning for libraries exceeding one billion compounds, while Isomorphic Labs says its IsoDDE can predict binding affinity and potential binding pockets from an amino-acid sequence.

Claude as the entry point

The product is generally available through Claude using Anthropic’s Model Context Protocol, or MCP, and through SandboxAQ’s website. MCP is a connection method that lets Claude call the specialized model, so a researcher can describe a screening task in plain language while SandboxAQ performs the quantitative analysis underneath.

SandboxAQ says a Google Cloud Marketplace release is planned next. The distribution plan puts the model into software environments researchers may already use, rather than requiring each team to build custom infrastructure around an inference endpoint. The scientific question, however, remains attached to the underlying model, its training data, and subsequent experimental validation.

The evidence still needed

SandboxAQ says AQPotency has produced experimentally validated impact in eight customer programs. It has not identified those programs or released model-wide benchmark results for accuracy, calibration, false positives, or prospective hit rates. That leaves outsiders unable to assess the model’s overall performance against other screening systems.

The launch therefore offers a clear technical proposition: screen more molecule-target pairs even when a solved protein structure is missing, at a claimed low marginal cost and with uncertainty flags attached. Whether that proposition reduces laboratory work or finds better starting points for drugs will be decided by predictions that survive the wet lab.